arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2601.22984cs.AI
收藏

为什么你的深度研究智能体会失败?关于完整研究轨迹中的幻觉评估

Why Your Deep Research Agent Fails? On Hallucination Evaluation in Full Research Trajectory

Zhejiang University(浙江大学);The University of Hong Kong(香港大学)

Yuhao Zhan, Tianyu Fan, Linxuan Huang, Zirui Guo, Chao Huang

AI总结:

针对深度研究智能体(DRA)在完整研究轨迹中累积的幻觉问题,提出从结果评估转向过程感知评估的PING分类法和细粒度评估框架,并构建DeepHalluBench基准,实验揭示系统性的可靠性差距。

摘要:

诊断深度研究智能体(DRA)的失败模式仍然是一个关键挑战。现有基准主要依赖端到端评估,掩盖了在研究轨迹中累积的中间幻觉。为弥补这一差距,我们提出从基于结果的评估转向过程感知评估,通过审计完整计划-搜索-总结轨迹中的幻觉。我们引入PING分类法,将DRA幻觉分为四种互补类型:传播、意图、噪声诱导和接地。我们进一步将该分类法实例化为一个细粒度评估框架,将轨迹分解为原子动作、声明和子查询以进行严格验证。利用该框架隔离100个特别容易产生幻觉的任务(包括对抗性场景),我们策划了DeepHalluBench。对六个代表性DRA的实验表明,在我们的幻觉压力测试集上,所有评估系统仍表现出不可忽视的可靠性差距。此外,我们的诊断分析将这些失败追溯到系统性缺陷,特别是幻觉传播和认知偏差,为未来的架构优化提供了可操作的见解。代码和数据可在https://github.com/yuhao-zhan/DeepHalluBench获取。

英文摘要:

Diagnosing failure patterns in Deep Research Agents (DRAs) remains a critical challenge. Existing benchmarks predominantly rely on end-to-end evaluation, obscuring intermediate hallucinations that accumulate throughout the research trajectory. To bridge this gap, we propose a shift from outcome-based to process-aware evaluation by auditing hallucinations in the full plan-search-summarize trajectory. We introduce the PING Taxonomy, which categorizes DRA hallucinations into four complementary types: Propagation, Intent, Noise-induced, and Grounding. We further instantiate this taxonomy into a fine-grained evaluation framework that decomposes trajectories into atomic actions, claims, and sub-queries for rigorous verification, and we validate its reliability on standard fact-checking benchmarks and human-reviewed trajectories. Leveraging this framework to isolate 100 hallucination-prone tasks, including adversarial scenarios, we curate DeepHalluBench. Experiments on six representative DRAs show that, on our hallucination-prone stress-test set, all evaluated systems still exhibit non-negligible reliability gaps. Furthermore, our diagnostic analysis traces these failures to systemic deficits, especially hallucination propagation and cognitive biases, providing actionable insights for future architectural optimization. Code and data are available at https://github.com/yuhao-zhan/DeepHalluBench.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑